272 Commits

Author	SHA1	Message	Date
xiaoxia	c7aed2c152	fix(#2180 P0): VLM/LLM timeout 调大，适配方舟视觉API 24-38s 响应 ... - VLM lite timeout: 15s → 45s（方舟VLM实测服务端处理24-38s，原15s必超时3次重试全挂） - VLM pro fallback timeout: 25s → 60s - 意图解析 chat timeout: 25s → 45s - 文案fast-first/fast-retry timeout: 25s → 45s - 文案pro-fallback timeout: 40s → 60s - 文案审核timeout: 20s → 30s - doubao_timeout默认: 30s → 45s 根因：方舟VLM视觉API从staging服务器实测响应24-38s(x-envoy-upstream-service-time:23357ms)， 原timeout=15s每次都被截断，3次重试全超时→lite降级pro又3次超时→总123s。 文案三级重试同理叠加到351s。调大timeout后单次调用应能在45s内返回。	2026-10-04 18:25:49 +08:00
CI Bot	07cf055a97	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-10-04 07:39:08 +00:00
xiaoxia	0e78f175fe	fix(#2174 P0): 信任链从i2i改为t2i文生图——方舟实测验证通过 ... #2170/#2172的信任链用Seedream图生图(i2i,传reference_images=用户真人照)， 实测产物不被Seedance信任，仍被400 portrait_intercept拦截。 方舟文档+实测确认：信任的是Seedream文生图(t2i)产物——纯prompt描述人物， 不传reference_images，产物是纯模型生成人像，同账号30天内可直接传Seedance。 改动： - VLM prompt新增portrait_prompt字段（60-100字人物外貌描述） - preheat_trust_chain签名从portrait_urls改为portrait_descriptions，t2i模式不传reference_images - video_generation移除现场跑i2i逻辑，仅使用预热好的t2i结果；预热失败走#2166自动降级t2v - 预热时机从VLM前改为VLM后（需要VLM输出的portrait_prompt），仍与文案阶段并行 - 单测适配：TestTrustChain全部改为pre_trusted_images路径验证 实测验证(充值后): - Seedream flash t2i ~15s出图，人像prompt中英文混合写实风格 - Seedance 2.5 i2v用t2i产物URL → 成功出5s视频(120s) - Seedance 2.0-mini i2v用t2i产物URL → 成功出5s视频 - 对照组直传用户真人照片 → 400 portrait_intercept（预期）	2026-10-04 15:31:38 +08:00
xiaoxia	371d1b9daf	fix(#2173 P0): 文案生成超时修复+VLM快速失败+Seedream换flash模型 (#2173) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-04 13:03:52 +08:00
xiaoxia	1a459d1ad2	perf(#2172): 信任链 Seedream i2i 与文案并行预热，缩短有参考图场景等待 (#2172) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-04 12:19:12 +08:00
saas-backend	8e19f24984	feat(#2170): 方舟信任链方案——真人照片经 Seedream AI 化后走 Seedance reference_image ... - 删除即梦 cvtob 接入代码（jimeng_client.py / JIMENG_* 配置 / _call_jimeng_video_generation） - 新增 DoubaoClient.image_generation：调用 Seedream 5.0 Pro (doubao-seedream-5-0-pro-260628) 文生图/图生图 - 信任链：有参考图时先经 Seedream 图生图 AI 化(保持五官特征)，AI 产物作为 reference_image 传 Seedance - 纯文字直传 Seedance 2.5；信任链强制走 reference_image 模式（非 first_frame），保留用户指定 ratio - Seedream 失败自动回退原图直传；图片/视频错误分类统一，支持 last_image_error - 删除 #2166 t2v 自动降级、#2169 jimeng 兜底逻辑 - config/base.py 新增 doubao_image_model/doubao_image_timeout，豆包模型升级到 seed-2-1 系列 - .env.example 清理 JIMENG_*，补全 VIDEO_*/IMAGE_* 配置 - points_rules.py 删除 jimeng-3.0 定价与模型配置 - 新增 10 个单测覆盖 Seedream 主路径 + 信任链 3 种场景；更新现有 mock	2026-10-04 02:44:08 +08:00
xiaoxia	866d71a431	feat(viral-video): #2169 接入即梦视觉API，真人参考图方舟50411自动切即梦兜底 (#2169) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-04 01:48:59 +08:00
saas-backend	c87810a4a6	fix(viral-video): 肖像/真人审核拦截时自动降级为纯文字t2v重试一次 ... - _step_render 捕获portrait_intercept/真人/肖像/隐私/SensitiveContent等关键词错误 - 第一次带参考图/首帧/参考视频被审核拦截时，自动移除所有参考素材走纯t2v重试 - 保留reference_audios(TTS口型)不影响肖像审核 - 仍失败则抛具体错误信息（#2166已增强的错误透传） - 解决用户上传真人产品图直接失败无法出片的问题	2026-10-03 23:25:58 +08:00
saas-backend	75ec9db439	fix(viral-video): #2165 视频生成错误详细日志+友好提示+真人拦截识别 ... - ai_client/dashscope_client 新增 last_video_error 字段，记录每次失败的 error_code/user_message/status_code/detail/model 等结构化信息 - 新增 _classify_video_error 错误分类器，自动识别： - 真人肖像拦截 (portrait_intercept) → 提示用户移除真人图片 - 鉴权失败 (auth_error) / 配额不足 (quota_exceeded) / 限流 (rate_limit) - 模型/Endpoint 不存在 (model_not_found) / 参数错误 (invalid_param) - 网络超时 (network_error) / 任务失败 (task_failed) / 未知 (unknown) - call_video_generation 失败时通过 get_last_video_error() 暴露详细错误 - worker _step_render 不再笼统抛"返回为空"，改为读取错误详情并展示具体用户提示 - 增强日志：完整 payload 打印（截断 prompt）、轮询 HTTP 错误记录完整 body、 下载/超时时记录具体失败原因 - 修复重试策略：4xx 错误不再重试（参数/鉴权错误重试无意义），仅 5xx/网络错误重试 - DashScope 客户端同样增加 last_video_error 和错误分类，保持语义一致 - 新增单测覆盖：400真人拦截、401鉴权失败、轮询failed、ai_service错误透传、dashscope 401	2026-10-03 23:25:58 +08:00
xiaoxia	1b02df4d4a	feat(viral-video): 多模型支持(4 Seedance+Wan 3.0)+DashScope接入+GET /models端点 (#2159) (#2161) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-03 17:42:38 +08:00
xiaoxia	76e11cb2f9	fix(viral-video) P0/P1: Seedance模型ID映射+很近全局替换+失败状态强落库 (#2157) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-03 16:51:25 +08:00
xiaoxia	6b2b30a6e9	chore(points): worker settle 打印预估/实际 tokens diff 便于公式校准 (#2155) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-03 02:03:14 +08:00
xiaoxia	5d6a4675fb	feat(viral-video): 动态积分定价（按tokens×单价×1.3，保留两位小数） (#2152) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-03 00:20:20 +08:00
saas-backend-agent	4a4b8f4a05	fix(viral-video): 错别字'很近'三保险修复 ... - 编导prompt加粗强调严禁使用'很近'，正确用词'最近' - _try_gen增加hard fix字符串替换：口播/分镜对白含'很近'自动替换 - 三级重试合格校验：口播含'很近'视为不合格触发重试 refactor(points): 移除ai_video/数字人/配音/改写/标题/封面/抖音解析扣点，仅保留声音克隆 - 删除5个路由的@points_gate装饰器(generation_preview/generation_cover/generation_tasks/scripts_ai/ai_avatar_render) - 删除tts.py/lipsync.py/narrative_service.py中ai_voice/ai_digital_human手动扣点+refund块 - points_rules.py精简POINTS_SCENES：仅保留voice_clone_train(0积分)和voice_clone_synth(1积分/分钟) - 删除DAILY_FREE_CLIP_LIMIT免费混剪额度逻辑，get_daily_usage返回unlimited占位 - points_gate.py删除ai_video特殊分支，未知/免费场景calculate_points_cost返回0直接放行 - points.py删除UNKNOWN_SCENE 400校验，任意scene可查 - 更新16个测试文件适配新规则：198个扣点相关单测全绿，viral_video 58个测试无回归	2026-10-02 20:39:16 +08:00
xiaoxia	70aeb5e642	fix(viral-video): P0 渲染卡死 + P1 VLM lite 快速降级（v1.6.1） ... P0 根因修复（v9 渲染卡死 20+ min）： 1. Seedance 2.5 content 数组协议错误：之前把参考音频/视频放在顶层 reference_audios/reference_videos（非官方字段）， 且首帧图不带 role 导致服务端识别为 first_frame 模式（该模式强制 ratio=adaptive），传入 ratio=9:16 导致任务 静默排队/失败轮询到 900s 超时。 2. 修复：所有参考（图/音/视）放进 content 数组并带正确 role（reference_image/reference_audio/reference_video）； 首张图根据是否有额外参考自动选择 first_frame vs reference_image 模式。 3. ratio 智能选择：纯首帧无参考→adaptive；omni_reference→按用户 ratio（默认 9:16）； 若创建任务因 ratio 返回 HTTP 400，自动回退 adaptive 重试一次。 4. 轮询 total_timeout 从 900s 收紧到 480s（8min），下载超时 120s； celery 任务加 soft_time_limit（render 660s/analyze 240s/generate_copy 180s/pipeline 900s）防挂死。 5. 轮询每 5 次打一次 info 日志，失败时完整打印 error.code/message。 6. _hb_stop 提前声明避免 finally UnboundLocalError。 P1 VLM lite 快速降级： - lite timeout 从 45s 收紧到 15s，pro 降级固定 25s，总耗时 ≤40s（之前 45+60≈100s）。 - 日志：创建任务时打印 mode（first_frame/omni_ref）、ratio、参考数量。	2026-10-02 16:34:56 +08:00
xiaoxia	d707b64876	tune(viral-video): VLM lite timeout 25s→45s (#2147) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-02 15:09:23 +08:00
CI Bot	0c0fe4619d	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-10-02 06:12:07 +00:00
xiaoxia	ad686dcd8b	fix(viral-video): P1 VLM模型ID/僵尸任务心跳/文案兜底/P2 ratio映射/OSS一致性 ... P1-1: VLM lite/pro模型ID修正（250915→250315/250328），根因是旧ID不存在导致lite必然超时降级pro - packages/config/base.py: doubao_vision_model/pro/lite 改为火山方舟官方正确ID - packages/shared/ai_client.py docstring同步更新 - apps/worker/worker_app/tasks/viral_video.py 硬编码字符串同步 - .env.example 同步新ID P1-2: worker心跳+僵尸任务超时回收 - packages/domain/viral_video.py: 新增heartbeat_at字段+touch_heartbeat()方法 - packages/adapters/sqlalchemy_impl/models.py: ViralVideoJobModel加heartbeat_at列(index=True) - packages/adapters/sqlalchemy_impl/viral_video_repository.py: to_entity/save/update映射heartbeat_at - alembic/versions/092_viral_video_heartbeat.py: 幂等迁移，加列+index+给已有running任务初始化 - apps/worker/worker_app/tasks/viral_video.py: - 新增_start_heartbeat_thread后台线程，每25s独立session更新heartbeat_at - 新增_recover_stale_jobs(): running>10min且心跳>2min未更新→自动标记failed - pipeline/analyze/generate_copy/render四个celery任务接入心跳启停+启动时回收僵尸 - apps/api/app/api/routes/viral_video.py: retry接口放宽，允许对超时僵尸任务重试 P1-3: 文案生成质量与兜底修复 - 兜底脚本'很近'→'最近'（之前commit已改） - _step_script_generation 三级重试: fast(0.8/2500tok)→fast(0.6/3200tok)→pro(0.7/3500tok) - 新增fallback判定：口播<20字/含兜底特征串/镜头数<1才视为降级，避免正常短脚本误判 - prompt关键要求新增'输出前自检：口播禁止错别字和语病' - 关键路径打info日志（模型/label/口播长度/镜头数/是否兜底），便于后续排查 P2-1: Seedance图生视频ratio不生效导致9:16出1:1 - 根因: 旧逻辑Bug #2110在image_url存在时不传ratio，模型默认adaptive→首帧方图商品图输出1:1 - 官方文档确认图生视频支持ratio参数，模型会自动居中裁剪首帧 - packages/shared/ai_client.py+ai_service.py: 删除image_url判空，始终传ratio - 同步更新docstring注释 P2-2: OSS成片eventual consistency导致completed后NoSuchKey - apps/worker/worker_app/tasks/viral_video.py: 新增_wait_oss_ready(url,timeout=10s) - mark_completed前循环HEAD公网URL，等200+content-length>0或10s超时 测试: viral-video相关97个单测全部通过，仅更新test_passes_reference_params_to_client以反映新ratio透传行为	2026-10-02 14:05:48 +08:00
xiaoxia	b31205e965	fix(viral-video): VLM lite失败自动降级pro + 10种IP人设注入prompt (#2139) (#2140) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-02 12:32:02 +08:00
xiaoxia	2a739dee17	fix(viral-video): #2134 generate-copy 提速 + 细粒度 phase/phase_message ... 问题7（generate-copy 提速，目标 30-40s）： - 新增 doubao_fast_model 配置（默认 doubao-1-5-pro-32k-250115），结构化输出任务（意图解析/编导脚本/合规审核）改用快模型，不再使用慢推理模型 doubao-seed-1-6 - call_llm 扩展支持 model/max_tokens/system_prompt 参数；chat_completion 同步支持 model 覆盖 - 编导脚本 temperature 0.8 + max_tokens 2500（从 4096 收紧）；意图解析 max_tokens 800；审核 max_tokens 500 - _SCRIPT_GENERATION_PROMPT 精简冗余描述（前导说明和关键要求章节从 ~70 行压到 ~30 行），减少输入/输出 token - 合规审核异步后置：阶段2 generate-copy 只做关键字黑名单快速检查（不调用 LLM），LLM 深度审核移到阶段3 confirm-copy TTS 之前执行，不再阻塞前端展示脚本 - 新增 _quick_compliance_blacklist_check 处理常见广告法绝对化用语 问题8（细粒度 phase + phase_message）： - ViralVideoJob 新增 phase_message 字段（中文提示文案，前端轮询直接展示） - SQLAlchemy ViralVideoJobModel 新增 current_stage/phase_message 列（current_stage 原已有但未持久化更新） - repo 层 _to_domain/save/update 同步处理新字段 - alembic 090 迁移：幂等 ADD COLUMN phase_message VARCHAR(500) - 新增 _set_stage 辅助：统一设置 current_stage + phase_message + Redis 推送 + DB 持久化 - 所有 celery task（analyze/generate-copy/render/one-click/resume）在关键节点调用 _set_stage 持久化阶段信息 - 阶段文案：analyzing_images→正在分析商品特征 / parsing_intent→正在解析文案意图 / generating_script→正在编排分镜脚本 / reviewing→合规审核中 / tts→正在合成AI配音 / rendering→正在生成视频 / uploading→正在上传视频 - ViralVideoJobResponse schema + _to_response 增加 current_stage/phase_message，前端轮询 GET /{job_id} 直接拿到 配套： - .env.example + render_env.sh SHARED_SECRETS 补 DOUBAO_FAST_MODEL - tests _FakeSettings/_make_job 同步新增字段 - _run_render_pipeline 兜底补生成分支不再同步调用 _step_review（由出片前统一审核处理）	2026-10-02 11:01:01 +08:00
xiaoxia	6b75eb5f67	feat(viral-video): VLM提速(lite+并行) + 商品描述增强(6维度+summary) + 爆款结构19选注入 (#2133) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-02 10:35:21 +08:00
xiaoxia	649420bd35	feat(viral-video): v1.6 单次Seedance出片+编导分镜脚本 #2124 ... - 输出从营销口播文案改为专业编导分镜脚本（CopyResult v1.6：overview/scene_and_lighting/shots/hard_constraints/negative_prompts/voiceover_script） - 新 SCRIPT_GENERATION 阶段替代原 copy_fusion+storyboard+review，prompt 指导 LLM 输出严格 JSON 结构 - pipeline 简化为4步：图片分析 → 编导脚本 → TTS整段合成(上传OSS做reference_audios) → 单次Seedance出片 → 上传 - 删除：多段分镜拆分、ffmpeg concat拼接、placeholder占位视频、分段重试降级、BGM单独混音（Seedance generate_audio=true原生合成音效/BGM） - ai_client/ai_service video_generation 支持 reference_images/reference_audios/reference_videos/generate_audio 参数 - duration 默认15秒，上限30秒；前端时长下拉 5/10/15/20/25/30s - 首帧图模式不传 ratio（保持 #2110 修复） - Alembic migration 088 幂等添加 voice_id/voice_source/video_ratio/video_model/copy_result 五列 - _step_image_analysis call_vision() 全路径 None 防护 - 向后兼容：final_copy=voiceover_script、storyboard=shots、老数据 _build_copy_result 降级拼装 - TS types 更新 CopyResult v1.6 结构 + ShotScript/Overview - 96个 viral 相关单测全通过	2026-10-01 13:55:31 +08:00
xiaoxia	069544da38	fix(viral-video): 三阶段 pipeline 补齐 voice_id/voice_source/video_ratio 字段 (#2123) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-01 13:05:38 +08:00
xiaoxia	d390d7c310	feat(viral-video): v1.5 三步分步流水线 analyze-images/generate-copy/confirm-copy (#2117) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-10-01 11:57:38 +08:00
CI Bot	a9cbe7d4c9	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-10-01 02:11:07 +00:00
xiaoxia	3a8ef857ac	fix(vlm): call_vision 走视觉模型而非文本模型 + 图片分析 prompt 结构化强化 ... 根因（VLM 牛头不对马嘴）：packages/shared/ai_service.call_vision() 之前误用 client.chat_completion（文本模型 doubao-seed-1.6-250615）发送多模态 content list。 文本模型不认图片 content type → 返回 None → _step_image_analysis 的 except 静默吞掉 → fallback 到占位结果 {name:'未识别'} → 后续 intent/copy/storyboard 完全没图信息，自然胡编。 修复： 1) call_vision 改用 client.vision_completion（模型 doubao-1-5-vision-pro-250915）， 走标准 OpenAI 多模态 chat/completions + image_url 格式 2) 超时提到 60s，温度降到 0.2，解析 ```json 代码块包裹 3) 增加详细 INFO 日志：打印 vision_model/image_url/prompt_len/原始返回前 400 字， 方便下次直接在 worker 日志排查 4) 重写 _IMAGE_ANALYSIS_PROMPT：强制结构化 JSON schema（category/name/brand/colors/ material_or_texture/key_features/visual_style/scene/target_audience_hint/ text_on_image），明确『无法判断就填无法判断，不许编造』，key_features 只能写外观可见特征、不许编功效 5) _step_image_analysis 健壮化：images 为空/None/文本返回/异常分别落 _source 标记； 每张图独立 try/except，单张失败不影响其他图 6) 下游 intent_parsing/copy_fusion/storyboard 兼容新字段 key_features/brand/ category/colors/visual_style（同时向后兼容旧 features 字段）	2026-10-01 10:06:59 +08:00
CI Bot	a74be7c717	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-09-30 20:43:44 +00:00
xiaoxia	cdce1b2e10	fix(viral-video): fix 6 E2E bugs — fusion_level alias, concat silent segments, ingest lookup, duplicated URL, TTS voice/format, Seedance first-frame ratio ... Bug1 (P0): schema accepts 'full_ai' as alias for 'ai_full' (Pydantic field_validator normalizes) Bug2 (P0): concat_video_files probes each segment audio stream; Seedance gen_audio=False segments now marked has_audio=False so concat filter uses aevalsrc silence instead of failing with ffmpeg exit 234 Bug3 (P1): find_by_storage_key now queries (storage_key OR file_url) to cover historical data where the legacy file_url column held assets/<project>/<date>/... paths Bug4 (P1): duplicated-hit response no longer accesses non-existent domain Asset.file_url; new helper _get_existing_asset_url uses storage_key (fallback file_url) through storage_service.get_url() Bug5 (P1): _step_tts passes job.persona_id as voice_id (default longxiaochun_v3) and forces format='mp3' so downstream ffmpeg -map 1:a:0 works regardless of provider Bug6 (P1): call_video_generation omits ratio param in first-frame (image_url) mode; ai_client.video_generation ratio becomes Optional[str] and is omitted from payload when None, fixing 400 InvalidParameter from Seedance	2026-10-01 04:10:06 +08:00
saas-bot	8bdc39a1ab	fix: P0 upload 404 (project/library mismatch) + worker bug fixes ... - upload.ts: library_id optional, auto ensureDefaultLibrary via getOrCreateDefaultProject - 5 call sites simplified (ViralVideoPage/useVoiceUpload/useBatchCovers/Step6Cover/useVoiceMaterials) - Bug1 (P0): upload_to_oss prefix kwarg → construct storage_key per generation.py pattern - Bug2 (P0): Seedance 404 enhanced logging w/ URL/model/response body + troubleshooting hints - Bug3 (P1): TTS import from services.* → apps.worker.services.* + synthesize fallback - duplicated=true: backend returns existing asset URL in prepare, frontend uses it - CI: ci_staging_deploy.sh auto-clean stale docker-compose.override.yml Refs: #2109	2026-09-30 22:14:28 +08:00
xiaoxia	22e04d65a7	refactor(viral-video): #2106 删除 _step_musetalk 步骤 ... 爆款视频由 Seedance 2.5 直接生成人物口型，不需要 MuseTalk 事后对口型。 MuseTalk 是 AI 数字人路线（上传人物视频+配音→对嘴型）用的，跟爆款视频是两条不同路线。 - 删除 _step_musetalk 函数 - resume_pipeline 直接把 render 输出传给 upload - 更新流水线 docstring 为 9 步 - 删除/更新对应单测 - ViralVideoStage.MUSETALK 枚举值保留以避免前端 breaking change	2026-09-30 20:42:12 +08:00
CI Bot	24724dca9f	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-09-30 19:06:33 +08:00
xiaoxia	d08835ec9f	fix(viral-video): #2106 P0 渲染阻塞修复 + Seedance 2.5 对接 + image_analysis 持久化 ... P0-1: Seedance 2.5 视频生成对接 - packages/shared/ai_client.py: DoubaoClient 新增 video_generation(prompt, image_url, duration, ratio, ...) 方法：走方舟 /contents/generations/tasks 异步任务（submit→poll→download），返回本地 MP4 路径 - packages/shared/ai_service.py: 新增 call_video_generation() 高层封装 - packages/config/base.py: 新增 doubao_video_model/doubao_video_timeout/doubao_video_poll_interval 配置 - apps/worker/.../viral_video.py _step_render 重写：按 storyboard 分镜逐段调 Seedance 生成短视频 → ffmpeg concat 拼接 → 混入 TTS 音频（-map 0:v/1:a -shortest -c:v copy） - 单分镜失败自动用 ffmpeg color 源占位片段兜底，保证 concat 不中断 - 新增 _normalize_storyboard/_fallback_storyboard/_build_segment_prompt/_probe_ok/_make_placeholder_clip 等辅助函数 P0-2: _step_video_analysis import 路径修复 - from worker_app.tasks.viral_video_analyzer → from viral_video.video_analyzer import analyze_video_style （文件在 apps/worker/viral_video/video_analyzer.py，worker PYTHONPATH 包含 apps/worker） - ImportError 仍兜底返回占位 style_guide，不阻塞流水线 P0-3: image_analysis 持久化 - packages/domain/viral_video.py: ViralVideoJob 新增 image_analysis: dict|None 字段 - packages/adapters/sqlalchemy_impl/models.py: viral_video_jobs 加 image_analysis JSON 列 - packages/adapters/sqlalchemy_impl/viral_video_repository.py: _to_domain/save/update 同步该字段 - alembic/versions/087_viral_video_image_analysis.py: migration 087 - run_viral_video_pipeline: step1 后立即 job.image_analysis = image_analysis 并 _save_job - resume_viral_video_pipeline: 从 job.image_analysis 读取，不再硬编码空 dict P1 顺手修复： - _step_tts: 返回值统一为 Path|None，Path 不存在/ImportError/synthesize 失败均返回 None - _step_bgm_select: BGM 素材未就绪前统一返回 None，渲染时跳过 BGM 混音 - _step_musetalk: GPU 端点未就绪前（即使有 persona_id）也直接跳过，不发 HTTP 请求 - TTS 工厂 import 路径修正: worker_app.services.tts_service_factory → services.tts_service_factory - credits_cost 赋值保留 TODO（等 credits.deduct() 总开关） - tests/unit/test_viral_video.py: BGM/TTS mock 对齐新返回语义 - tests/unit/test_viral_video_p0.py: 新增 15 个单测覆盖 video_analysis/storyboard 规范化/ TTS Path 处理/BGM None/MuseTalk 跳过/call_video_generation 委托/placeholder clip/resume 读 job	2026-09-30 19:06:33 +08:00
xiaoxia	ca834b23cb	feat(api+worker): 爆款视频 WebSocket 实时进度推送端点（#2051） (#2103) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-30 10:44:29 +08:00
xiaoxia	34305974ad	Merge branch 'pr2101' into merge-test ... # Conflicts: # apps/api/app/schemas/viral_video.py # apps/worker/worker_app/tasks/viral_video.py # packages/adapters/sqlalchemy_impl/viral_video_repository.py # packages/domain/viral_video.py # tests/unit/test_viral_video.py	2026-09-30 08:01:26 +08:00
xiaoxia	c45a2ce9b1	fix(worker): viral_video Celery 任务改用 @shared_task，修复单测在 CI 下 resume_viral_video_pipeline.run 找不到的问题 ... - @celery_app.task 在 conftest mock Task.delay 之后若模块重新加载，decorator 返回的是未绑定原函数而非 Task 实例，导致 .run 方法不存在 - @shared_task（和 lipsync_tts.py 保持一致）在任务被任意 Celery app 加载时自动绑定，测试和生产环境行为一致 - 保留 celery_app import 为 noqa F401，确保 worker_app.celery_app 模块加载时会注册 viral_video tasks	2026-09-30 00:30:38 +08:00
xiaoxia	6636dc45f7	feat(#2039): viral video domain + repository + REST API + Celery orchestrator（PR2/2） ... - domain: ViralVideoJob 状态机（PENDING/RUNNING/WAIT_USER_CONFIRM/COMPLETED/FAILED/CANCELLED）、 11 阶段枚举、PromptType 7 值（含 v1.3 video_style_integration/style_constraint） - repository: 接口 + SQLAlchemy 实现（jobs/style_templates/prompt_templates） - API: 6 个 REST 端点（generate/history/detail/retry/confirm-intent/analyze-style/style-templates） - Celery: ViralVideoOrchestrator 10 步流水线，WS viral_video:progress 进度推送， Credits CREDITS_VIRAL_VIDEO_COST=50 扣点/失败自动回滚 - ai_service: 新增 call_llm/call_vision（复用现有豆包客户端） - 测试 40 个单测（domain/schema/repository/流水线/集成）	2026-09-29 22:22:04 +08:00
CI Bot	a7d6ba473b	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-09-29 13:12:48 +00:00
xiaoxia	f19be5fd09	feat(#2039): viral video domain + repository + REST API + Celery orchestrator（PR2/2） ... - domain: ViralVideoJob 状态机（PENDING/RUNNING/WAIT_USER_CONFIRM/COMPLETED/FAILED/CANCELLED）、 11 阶段枚举、PromptType 7 值（含 v1.3 video_style_integration/style_constraint） - repository: 接口 + SQLAlchemy 实现（jobs/style_templates/prompt_templates） - API: 6 个 REST 端点（generate/history/detail/retry/confirm-intent/analyze-style/style-templates） - Celery: ViralVideoOrchestrator 10 步流水线，WS viral_video:progress 进度推送， Credits CREDITS_VIRAL_VIDEO_COST=50 扣点/失败自动回滚 - ai_service: 新增 call_llm/call_vision（复用现有豆包客户端） - 测试 40 个单测（domain/schema/repository/流水线/集成）	2026-09-29 19:41:51 +08:00
saas-backend-agent	96bcec5fdd	fix(worker+api): resolve batch title race + auto-queue pending tasks (#2098) ... Two batch-rendering bugs found from staging logs (6-video batch 17:30-17:43): Bug A (P1) — title_config race when same plan renders concurrently: _sync_task_config_to_plan wrote each task's title_config/bgm/resolution to the shared plan.config row, then RenderAdapter.render_plan read back from plan.config during rendering. If two tasks sharing the same plan (e.g. batch retry, preview+render, retry+new) ran concurrently, Task B's write could overwrite Task A's title before Task A's ffmpeg read it, producing videos with the wrong title text/style. Fix: stop mutating plan.config in the worker render path. Introduce task_config_override threading through render_plan / _do_render / UnifiedRenderService; URS reads config via _effective_config() which does a deep-copy plan.config merged with per-task override (title/bgm/ export). _build_task_config_override builds the override dict (with field-name normalization: font_size→size, font_color→color) from task_info, and _download_voice_for_task handles voiceover download independently. plan.config is left as-authored by the API writeback; concurrent tasks no longer race on it. Files: - apps/worker/video_processing/unified_render_service.py: add override_config param + _cfg_section/_effective_config helpers; replace all render-time self.plan.config reads (title/subtitle/bgm/ export/TTS) with _effective_config(). - apps/worker/video_processing/render_adapter.py: render_plan/_do_render/ _prepare_bgm accept task_config_override, merge into bgm/export reads, forward to URS. - apps/worker/worker_app/tasks/generation.py: replace _sync_task_config_to_plan (which wrote plan.config) with _build_task_config_override + _download_voice_for_task; pass task_config_override to render_plan. Bug B (P0) — USER_PENDING_LIMIT=3 returns HTTP 429, blocks batch submit: Pre-check in create_generation_task rejected the whole batch with 429 USER_QUEUE_FULL once user had ≥3 pending tasks; UI showed 'wait ~4 min' and prevented any more submissions. Users expect to submit a batch of 6 and have them queue naturally (worker concurrency=2). Fix: - USER_PENDING_LIMIT 3→20 (soft cap for abuse protection, supports typical batch sizes of 6-10 with headroom). - Remove user-level 429 rejection from pre-check, retry endpoint, single-task confirm endpoint, and safe_enqueue (user path now logs a warning and continues to enqueue). Global GLOBAL_PENDING_LIMIT=20 is retained as a hard 503 system-busy guard. - safe_enqueue post-enqueue check: user-over only logs, does not fail the task or raise. - Batch loop UserPendingLimitExceeded except branch now treats it as a successful enqueue (should not trigger in practice). Files: - apps/api/app/core/task_enqueue.py: limit 3→20; user checks log-only. - apps/api/app/api/routes/generation_tasks.py: remove user pre-check 429; soften retry/single/batch except branches. Stacked on PR #2095/#2097 which already fixed title baseline/width-scaling and default-constant alignment with CPU video_filter_builder. Tests: - tests/unit/test_gpu_direct_pipeline.py: 30 passed - tests/unit/test_render_adapter_pure.py: 18 passed - ruff check/format clean; py_compile clean	2026-09-29 18:31:05 +08:00
xiaoxia	3106496c12	fix(worker): GPU直连管线音频bug修复+边缘裁剪入GPU省5s (#2089) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-28 20:17:54 +08:00
Agent	4d7c80ae07	fix: OSS签名URL校验HEAD→GET+Range，修复403误报 ... OSS sign_url()将路径中/编码为%2F，HEAD请求返回403但GET正常。 改用GET+Range:bytes=0-0仅下载1字节验证，同时接受206状态码。 根因修复，消除URL校验失败后的object_exists兜底路径。	2026-09-28 18:01:12 +08:00
xiaoxia	1a93a9c00e	fix(quality): classification_status 赋枚举值而非裸字符串 + 仓储层防御 (#2079) (#2080) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-28 01:37:20 +08:00
xiaoxia	59c05148ab	fix(p0): finalize 400（extra_meta漏传）+ 查重失效（fingerprint_chunks漏存） (#2068) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-27 20:47:12 +08:00
CI Bot	39c064ba1c	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-09-26 20:43:21 +08:00
saas-backend-agent	105ab54059	feat(gpu): 接入 P4000 NVENC 硬件编码加速 ... - packages/config/base.py: 新增 GPU_ENCODE_* 配置项（开关/endpoint/relay/secret/编码参数/fallback） - packages/shared/gpu_encoder.py: 新增 GpuEncoderClient，封装 health 探测 + mezzanine 上传 OSS + P4000 nvenc 编码 + relay 回传下载，失败抛 GpuEncodeError 触发 CPU 降级 - apps/api/app/api/routes/gpu_relay.py: 新增 internal PUT/GET/DELETE /api/v1/internal/gpu-relay/{key}（token 鉴权），P4000 PUT 编码结果，worker GET 下载 - apps/api/app/api/router.py: 注册 gpu_relay_router - apps/worker/video_processing/unified_render_service.py: _execute_ffmpeg 和 _render_pass_through 尝试 GPU 路径：CPU ultrafast mezzanine → P4000 nvenc → 输出到最终路径；任何失败自动回退到原 CPU libx264 路径 - apps/worker/worker_app/tasks/_startup.py: worker_ready 时探测 P4000 健康并打日志 - tests/unit/test_gpu_encoder.py: GpuEncoderClient 单测（health/sync 调用/失败/fallback/relay URL） 架构： - 输入：CPU 输出 libx264 ultrafast mezzanine → 上传 OSS 临时前缀 → P4000 签名 URL 下载 - 输出：P4000 PUT → 宿主机 nginx（tailscale:80）→ API /api/ 反代 → gpu_relay 路由落盘到 generated/gpu_relay/ - 回传：worker 通过 docker 网络 http://xiaoxia-api-staging:8000 GET 下载最终 mp4 到 output_path - 降级：GPU 任何环节异常（health/上传/编码/回传/下载）→ 原 CPU 路径继续执行，不影响成片	2026-09-26 20:43:21 +08:00
xiaoxia	8a396303b8	fix(voice-clone): 处理中卡死兜底 — worker重启/Celery消息丢失后processing永久卡住 (#2050) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-26 15:31:06 +08:00
xiaoxia	751f8ad84e	feat(#2035): 语义标签增强 + 质量评分自动计算 + atom_clip caption/embedding (#2036) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-25 12:44:37 +08:00
xiaoxia	23fe5f9822	feat(generation): #2024 视频渲染后延迟到封面确认才入库（新增 AWAITING_COVER 状态 + finalize 接口） (#2026) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-24 12:34:43 +08:00
xiaoxia	38ffa0b98b	fix(1978): register lipsync_gpu task in worker celery imports ... Worker celery_app.conf.imports was missing app.tasks.lipsync_gpu, so lipsync_gpu_process_async.apply_async() messages were sent to broker but never consumed by any worker, leaving lipsync_jobs stuck at status=processing forever.	2026-09-20 13:44:10 +08:00
xiaoxia	a1f25a4426	feat(worker): #1970 AI 标签 backfill 支持 force 重打降级记录 (#1989) ... Co-authored-by: xiaoxia <dev@xiaoxiajianji.com> Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>	2026-09-19 10:49:21 +08:00